Year 2087: a golden glass beehive megastructure orbits above the Bay of Bengal at sunrise; its cells are packed with glowing golden bits that pour down toward the Ganges delta while small drones circle it like bees.

TensorHive · measured on 2× NVIDIA DGX Spark

Frontier models.
Your desk.
Written in Rust.

TensorHive is a memory-safe LLM inference engine for NVIDIA's desktop Blackwell systems, Rust from the server down to the GPU kernels. DGX Spark today, RTX PRO next.

Get early access See the benchmarks →

For teams running NVIDIA DGX Spark or RTX PRO who want frontier models without the cloud.

2×DGX Spark (GB10)

256 GBunified memory

~6 µs16 KiB exchange, p50

Rusthost, transport & GPU kernels

Measured, not promised

Every number here comes from a run on our two DGX Sparks that wrote its own evidence file.

All benchmarks & receipts →

321B

Served on two Sparks

GLM-5.3-Flash, tensor-parallel. 16 / 16 replies match the reference run over 2,032 steps.

~6 µs

Rank-to-rank exchange

Median for a 16 KiB exchange over ConnectX-7, GPU in the loop, on our own RDMA transport.

89%

Of GB10 memory bandwidth

Rust GPU kernels at 240–245 GB/s of the 273 GB/s peak, bit-exact against a CPU reference.

TensorHive

Built on NVIDIA's cuda-oxide Rust compiler. One toolchain, one binary, no Python in the serving path.

Memory-safe end to end

Server, scheduler, network transport and GPU kernels in Rust. The last few borrowed CUDA kernels are being replaced one by one.

Built for GB10

Kernels tuned to DGX Spark's unified memory, with our own RDMA transport for multi-node serving.

Verified, not assumed

Outputs checked step by step against a reference run. A change lands only with evidence.

Private by design

Runs on your machines behind an OpenAI-compatible API. Your data stays with you.

Today

NVIDIA DGX Spark

Chip: GB10 Grace Blackwell
Memory: 128 GB unified, 273 GB/s
Scale-out: ConnectX-7, 200 Gb/s

Memory: 96 GB GDDR7
Bandwidth: ~1.8 TB/s
Why: Same Blackwell family, more speed

Run large models on hardware you own

Using DGX Spark or RTX PRO and want private inference? Join the early-access list, or write to contact@mouchak.ai.

Get early access

Frontier models.Your desk.Written in Rust.